Re: excess precision for 'float'

[Bastien ROUCARIES]

On Sun, Oct 9, 2011 at 12:31 AM, Bruno Haible <address@hidden> wrote:
> It is well-known that 'double' computation results on x86 processors can
> carry excess precision. Namely, if the value is stored in a floating-point
> register (80 bits total, 64 bits mantissa) without passing through a
> memory store, it can carry excess precision. The two ways to prevent
> such excess precision (with GCC) are:
>  - option -ffloat-store,
>  - cast to 'volatile double'.
>
> It is not so well-known that this phenomenon can also happen with 'float'
> computations.
>
> Here's a program:
> ======================================================================
> #include <stdio.h>
> #include <math.h>
> int main ()
> {
>  printf ("%.12g\n", (double) 0.9272951f);
>  printf ("%.12g\n", (double) 0.92729513f);
>  printf ("%.12g\n", (double) 0.92729514f);
>  printf ("%.12g\n", (double) 0.92729515f);
>  printf ("%.12g\n", (double) 0.92729516f);
>  printf ("%.12g\n", (double) 0.92729517f);
>  printf ("%.12g\n", (double) 0.92729518f);
>  printf ("%.12g\n", (double) 0.92729519f);
>  printf ("%.12g\n", (double) 0.9272952f);
>  printf ("%.12g\n", (double) 0.9272953f);
>  printf ("\n");
>  printf ("%.12g\n", (double) atan2f (0.8f, 0.6f));
>  printf ("%.12g\n", (double) (float) atan2f (0.8f, 0.6f));
>  printf ("%.12g\n", (double) (volatile float) atan2f (0.8f, 0.6f));
>  return 0;
> }
> ======================================================================
>
> Output on mingw:
>
> 0.927295088768
> 0.927295148373
> 0.927295148373
> 0.927295148373
> 0.927295148373
> 0.927295148373
> 0.927295207977
> 0.927295207977
> 0.927295207977
> 0.927295327187
>
> 0.927295206081
> 0.927295206081
> 0.927295206081
>
> You can see that there are no 'float' values between
> 0.927295148373 and 0.927295207977, yet the atan2f() function
> - which is declared to return 'float' - has returned a value
> between them.
>
> Does anyone know whether this can also happen with glibc functions?
> Or are glibc functions that return 'float' really returning 'float'
> values?

This can happen and it is clearly a bug for me (as a hpc coder) but
not for gcc maintener.

Serge L of inria is the person to contact for this case. Will find his contact.

Thanks

Bastien


> Bruno
> --
> In memoriam Engin Ceber <http://en.wikipedia.org/wiki/Engin_Ceber>
>
>